A Transformer-Based Deep Learning Approach to Predicting Air Organic Pollutant-Human Protein Interactions.

Area of Science:

• Environmental Health
• Computational Biology
• Toxicology

Background:

• Air pollution is a major global health concern.
• Understanding molecular interactions between pollutants and proteins is crucial for assessing toxicity.
• Current methods struggle to identify these interactions, especially for new pollutants.

Purpose of the Study:

• To develop a novel deep learning model for predicting airborne organic pollutant-protein interactions.
• To enhance mechanistic understanding and risk assessment of air pollution toxicity.

Main Methods:

• Developed tipFormer, a deep learning model using dual pretrained language models and cross-attention.
• Encoded proteins and organic pollutants to capture interaction patterns.
• Validated predictions using genome-wide transcriptomic analysis in human bronchial epithelial cells.

Main Results:

• tipFormer achieved state-of-the-art performance with an AUC of 0.9787 on a test set.
• Predicted pollutant targets showed significant concordance with experimentally responsive genes.
• Demonstrated biological relevance and mechanistic insight into air pollution effects.

Conclusions:

• tipFormer offers a powerful computational approach for predicting pollutant-protein interactions.
• The study provides deeper mechanistic insights into the molecular basis of air pollution toxicity.
• This work bridges computational predictions with experimental validation for improved risk assessment.